| <!doctype html> |
| <html lang="en"> |
| <head> |
| <meta charset="utf-8" /> |
| <meta name="viewport" content="width=device-width, initial-scale=1" /> |
| <meta name="theme-color" content="#0c1924" /> |
| <meta name="description" content="Explore SAEBase: activation storage, feature interventions, benchmark measurements and reliability checks across 48 verified model/SAE layer sites." /> |
| <title>SAEBase Research Lab Β· OpenSAE</title> |
| <link rel="stylesheet" href="styles.css" /> |
| </head> |
| <body> |
| <div class="shell"> |
| <aside class="sidebar" aria-label="Primary navigation"> |
| <a class="brand" href="https://huggingface.co/OpenSAE" target="_blank" rel="noopener noreferrer" aria-label="OpenSAE organization"> |
| <span class="brand-glyph" aria-hidden="true"><svg viewBox="0 0 48 48"><g fill="none" stroke="currentColor" stroke-width="2.3" stroke-linecap="round"><path d="M24 24V11"/><path d="M24 24V11" transform="rotate(60 24 24)"/><path d="M24 24V11" transform="rotate(120 24 24)"/><path d="M24 24V11" transform="rotate(180 24 24)"/><path d="M24 24V11" transform="rotate(240 24 24)"/><path d="M24 24V11" transform="rotate(300 24 24)"/><circle cx="24" cy="7" r="2.5"/><circle cx="24" cy="7" r="2.5" transform="rotate(60 24 24)"/><circle cx="24" cy="7" r="2.5" transform="rotate(120 24 24)"/><circle cx="24" cy="7" r="2.5" transform="rotate(180 24 24)"/><circle cx="24" cy="7" r="2.5" transform="rotate(240 24 24)"/><circle cx="24" cy="7" r="2.5" transform="rotate(300 24 24)"/><circle cx="24" cy="24" r="4"/></g></svg></span> |
| <span>OpenSAE<small>Research infrastructure</small></span> |
| </a> |
| <div class="side-label">WORKSPACE</div> |
| <nav id="nav"> |
| <button class="nav-item active" data-tab="overview"><span>β«</span> Overview</button> |
| <button class="nav-item" data-tab="tutorial"><span>β·</span> 5-minute tutorial</button> |
| <button class="nav-item" data-tab="explorer"><span>β¦</span> Verified checkpoints</button> |
| <button class="nav-item" data-tab="measurements"><span>β</span> Measurement explorer</button> |
| <button class="nav-item" data-tab="workflow"><span>β</span> Research workflow</button> |
| <button class="nav-item" data-tab="getstarted"><span>β</span> Get started</button> |
| </nav> |
| <div class="side-bottom"> |
| <span class="live-dot"></span><span>Release v0.3.0</span> |
| <p>Static explorer Β· frozen source records</p> |
| <a href="https://github.com/OpenSAE/SAEBase" target="_blank" rel="noopener">View GitHub repository β</a> |
| </div> |
| </aside> |
| <main class="main"> |
| <header class="topbar"><span>OPENSAE <span class="crumb">/</span> SAEBASE <span class="crumb">/</span> <span id="crumb">OVERVIEW</span></span><a class="toplink" href="https://github.com/OpenSAE/SAEBase/releases/tag/v0.3.0" target="_blank" rel="noopener">Software release <span>β</span></a></header> |
| <section class="page active" data-page="overview"> |
| <div class="hero"> |
| <div class="hero-copy"><div class="eyebrow"><span class="eyebrow-dot"></span> RELEASE v0.3.0 Β· OPEN RESEARCH INFRASTRUCTURE</div><h1>Study the features.<br><em>Measure the behavior.</em></h1><p>SAEBase connects activation capture, feature interventions, quantitative benchmarks, and reliability analysis in one reproducible research workflow.</p><div class="action-row"><button class="primary" data-go="tutorial">Try the 5-minute tutorial <span>β</span></button><button class="button-ghost" data-go="explorer">Explore verified checkpoints β</button></div></div> |
| <div class="hero-art" aria-hidden="true"><svg viewBox="0 0 420 310" role="img"><defs><linearGradient id="wire" x1="0" y1="0" x2="1" y2="1"><stop stop-color="#66ebcc"/><stop offset="1" stop-color="#62aeea"/></linearGradient></defs><g opacity=".14" stroke="#c0f2e6" stroke-width="1"><path d="M0 62H420 M0 124H420 M0 186H420 M0 248H420 M70 0V310 M140 0V310 M210 0V310 M280 0V310 M350 0V310"/></g><g stroke="url(#wire)" stroke-width="2" fill="none" opacity=".85"><path d="M38 152L116 86L207 149L298 80L380 147 M38 152L114 214L207 149L301 217L380 147 M116 86L114 214 M298 80L301 217 M207 149L207 34 M207 149L207 278"/></g><g fill="#0d2f3f" stroke="#7ce7cf" stroke-width="2.2"><circle cx="38" cy="152" r="11"/><circle cx="116" cy="86" r="13"/><circle cx="114" cy="214" r="11"/><circle cx="298" cy="80" r="13"/><circle cx="301" cy="217" r="11"/><circle cx="380" cy="147" r="11"/><circle cx="207" cy="34" r="9"/><circle cx="207" cy="278" r="9"/></g><circle cx="207" cy="149" r="34" fill="#124b57" stroke="#78edcb" stroke-width="2.7"/><circle cx="207" cy="149" r="15" fill="#b6ffea" opacity=".95"/><circle cx="207" cy="149" r="56" fill="none" stroke="#77dbc7" stroke-dasharray="4 9" opacity=".45"/></svg><div class="art-label">MODEL REPRS. <span>β</span> FEATURE INTERVENTIONS</div></div> |
| </div> |
| <div class="kpis"><div class="kpi"><div>MODELβSAE CONFIGURATIONS</div><strong id="overviewProfiles">18</strong><small>across 270Mβ9B parameter models</small></div><div class="kpi"><div>VERIFIED MODEL/LAYER SITES</div><strong id="overviewSites">48</strong><small>real checkpoint compatibility</small></div><div class="kpi"><div>BENCHMARK + RELIABILITY</div><strong>48 / 48</strong><small>recorded evaluations and audits</small></div><div class="kpi"><div>PUBLIC SOFTWARE</div><strong>v0.3.0</strong><small>GitHub Release Β· Apache-2.0</small></div></div> |
| <div class="section-heading"><div><div class="kicker">CAPABILITIES</div><h2>One workflow, from representations to evidence</h2></div></div> |
| <div class="feature-grid"><article class="feature"><div class="symbol">β£</div><h3>Activation store</h3><p>Capture once, reuse across SAE checkpoints and experiments with identity-aware caching.</p></article><article class="feature"><div class="symbol">β</div><h3>Feature interventions</h3><p>Ablation, scaling, clamping, coalition edits and downstream behavioral measurements.</p></article><article class="feature"><div class="symbol">β₯</div><h3>Benchmarking</h3><p>Reconstruction, sparsity and intervention measurements with frozen experiment configurations.</p></article><article class="feature"><div class="symbol">β</div><h3>Reliability</h3><p>Group-aware statistical auditing, source provenance, persistent reports and repeatability checks.</p></article></div> |
| <div class="note"><span class="note-icon">i</span><span><strong>Traceable evaluation.</strong> Browse each checkpoint, inspect its recorded measurements, and open the original source report. Results follow a controlled evaluation protocol; they are not a cross-model quality ranking.</span></div> |
| </section> |
| <section class="page" data-page="explorer"> |
| <div class="page-head"><div class="kicker">VALIDATED CHECKPOINTS</div><h1>Verified model & SAE coverage</h1><p>Browse 48 real-weight model/layer sites across 18 SAE configurations. Select any row to inspect the exact hook, metrics, and original verification report.</p></div> |
| <div class="kpis three"><div class="kpi"><div>SELECTED SITES</div><strong>48</strong><small>all passed compatibility checks</small></div><div class="kpi"><div>CONFIGURATIONS</div><strong>18</strong><small>TopK Β· JumpReLU Β· variants</small></div><div class="kpi"><div>HELD-OUT EVALUATION</div><strong>64</strong><small>controlled prompt groups per site</small></div></div> |
| <div class="work-panel"><div class="panel-top"><div><h2>Checkpoint browser</h2><p>Filter by checkpoint, architecture or model family.</p></div><div class="record-count" id="recordCount">48 records</div></div><div class="filters"><input id="searchBox" type="search" placeholder="Search model, SAE or hook β¦" aria-label="Search checkpoint records"/><select id="familySelect" aria-label="Filter model family"><option value="all">All families</option></select><select id="archSelect" aria-label="Filter SAE architecture"><option value="all">All architectures</option></select></div><div class="table-scroll"><table aria-label="Verified model and SAE sites"><thead><tr><th>MODEL</th><th>SAE ARCHITECTURE</th><th>LAYER</th><th>COMPATIBILITY</th><th>RESULT</th></tr></thead><tbody id="siteRows"></tbody></table></div></div> |
| <div class="detail-card" id="siteDetail"><div class="detail-empty">Select a checkpoint row to inspect its measurements and source report.</div></div> |
| <div class="note"><span class="note-icon">i</span><span>These values refer to selected model/SAE hooks, not every possible checkpoint. The original source contains weights provenance, report identities, and independent reliability auditing.</span></div> |
| </section> |
| <section class="page" data-page="measurements"><div class="page-head"><div class="kicker">ORIGINAL EVALUATION RECORDS</div><h1>Explore recorded measurements</h1><p>Inspect precomputed numeric results from genuine checkpoint runs. No model execution happens in this browser.</p></div> |
| <div class="work-panel"><div class="panel-top"><div><h2>Within-checkpoint layer profile</h2><p>Choose a model and a metric; comparisons are confined to its recorded layers.</p></div></div><div class="filters"><select id="metricModel" aria-label="Choose model and SAE"></select><select id="metricField" aria-label="Choose metric"><option value="explained_variance">Explained variance</option><option value="mean_l0">Feature count (L0)</option><option value="reconstruction_mse">Reconstruction MSE</option><option value="feature_ablation_logit_delta_l2">Feature-ablation logit Ξ L2</option></select></div><div class="chart-box"><svg viewBox="0 0 730 280" id="metricChart" role="img" aria-label="Recorded values over model layers"></svg></div><p class="caption">Values are recorded engineering measurements on 64 held-out controlled prompt groups per site. Reconstruction scales and readout units may differ across checkpoints; do not compare model families as a quality ranking.</p></div> |
| <div class="section-heading"><div><div class="kicker">INSPECT THE SOURCE</div><h2>Follow every measurement back to the experiment</h2></div></div><div class="link-grid"><a href="https://github.com/OpenSAE/SAEBase/blob/main/docs/verified-model-matrix.md" target="_blank" rel="noopener"><strong>Verification matrix</strong><span>Per-layer report links β</span></a><a href="https://github.com/OpenSAE/SAEBase/releases/tag/stage-d-original-raw-20261016" target="_blank" rel="noopener"><strong>Original run archives</strong><span>Immutable source measurements β</span></a><a href="https://huggingface.co/datasets/OpenSAE/SAEBase-Evaluation" target="_blank" rel="noopener"><strong>Evaluation dataset</strong><span>Machine-readable measurement rows β</span></a></div> |
| </section> |
| <section class="page" data-page="workflow"><div class="page-head"><div class="kicker">RESEARCH WORKFLOW</div><h1>From activation to reliable evidence</h1><p>A single framework for collecting SAE representations, measuring interventions and reporting results.</p></div><div class="pipeline"><div><small>01 Β· INPUTS</small><h3>Model & SAE</h3><p>Identify source checkpoints, exact layers and activation hooks.</p></div><span>β</span><div><small>02 Β· CACHE</small><h3>Activation store</h3><p>Capture and reuse model activations and sparse feature codes.</p></div><span>β</span><div><small>03 Β· EDIT</small><h3>Feature interventions</h3><p>Modify selected features and measure downstream effects.</p></div><span>β</span><div><small>04 Β· AUDIT</small><h3>Benchmarks & reliability</h3><p>Produce versioned metrics and statistically audited reports.</p></div></div> |
| <div class="section-heading"><div><div class="kicker">RESEARCH INTEGRATIONS</div><h2>Tools built on the same foundations</h2></div></div> |
| <div class="research-grid"><a href="https://github.com/OpenSAE/SAE-Behavioral-Compositionality" target="_blank" rel="noopener"><small>FEATURE INTERACTIONS</small><h3>SAR</h3><p>Measure non-additive downstream effects between SAE features.</p><span>Explore research β</span></a><a href="https://github.com/OpenSAE/Redistribution-Sensitivity-Score" target="_blank" rel="noopener"><small>PROMPT TRANSFORMATIONS</small><h3>RSS</h3><p>Study consistent feature redistribution across input presentations.</p><span>Explore research β</span></a><a href="https://github.com/OpenSAE/SAE-Input-Localization" target="_blank" rel="noopener"><small>INPUT DEPENDENCIES</small><h3>Input Localization</h3><p>Study the inputs needed to explain learned feature activations.</p><span>Explore research β</span></a></div> |
| <div class="note"><span class="note-icon">i</span><span>Research estimators remain in their own repositories; SAEBase supplies common capture, intervention, evaluation and provenance infrastructure.</span></div> |
| </section> |
| <section class="page" data-page="tutorial"><div class="page-head"><div class="kicker">5-MINUTE TOUR</div><h1>Your first reproducible SAE experiment</h1><p>Follow the key stages of a SAEBase run. This walkthrough explains the real software workflow using examples and links; no model runs in the browser.</p></div> |
| <div class="tour-layout"><div class="tour-side"><div class="kicker">THE WORKFLOW</div><div id="tourNav"></div><div class="tour-side-foot">6 short steps Β· Python 3.11β3.12 Β· start without GPU</div></div><div class="tour-content"><div class="tour-step-heading"><div class="kicker" id="tourStepNo">STEP 01 / 06</div><button class="copy" id="copyTourCode" type="button">Copy command</button></div><h2 id="tourTitle">Start with a known environment</h2><p id="tourDescription"></p><div class="tour-illustration"><div class="tour-small" id="tourConcept">ACTIVATION β SAE β INTERVENTION β MEASUREMENT</div><div class="tour-pipeline" id="tourPipeline"></div></div><pre id="tourCode"><code></code></pre><div class="tour-actions"><button id="tourPrev" class="secondary-action">β Previous</button><span id="tourProgress">1 of 6</span><button id="tourNext" class="primary">Next step β</button></div></div></div><div class="note"><span class="note-icon">i</span><span>Use the linked GitHub tutorial to execute the experiment locally. The browser tour explains which files are generated and how they are verified; it does not substitute synthetic values for model evidence.</span></div></section> |
| <section class="page" data-page="getstarted"><div class="page-head"><div class="kicker">FOR DEVELOPERS & RESEARCHERS</div><h1>Build with SAEBase</h1><p>Begin with an offline reproducible experiment, then bring your own model, SAE or benchmark metric.</p></div><div class="work-panel"><div class="panel-top"><div><h2>Five-minute local quickstart</h2><p>Python 3.11β3.12 Β· no pretrained checkpoint or GPU required for this first example.</p></div><button class="copy" id="copyCode">Copy code β</button></div><pre id="installCode"><code>git clone https://github.com/OpenSAE/SAEBase.git |
| cd SAEBase |
| python -m pip install -e . |
| python examples/quickstart/first_experiment.py --output runs/first-experiment</code></pre><p class="caption">The quickstart follows the current GitHub main branch and uses a deterministic toy model. The dashboard measurements remain pinned to the independently published v0.3.0 evaluation records.</p></div><div class="link-grid"><a href="https://github.com/OpenSAE/SAEBase" target="_blank" rel="noopener"><strong>Source code</strong><span>GitHub β</span></a><a href="https://github.com/OpenSAE/SAEBase/releases/tag/v0.3.0" target="_blank" rel="noopener"><strong>Installable release</strong><span>Wheel and source archive β</span></a><a href="https://github.com/OpenSAE/SAEBase/blob/main/docs/extensions.md" target="_blank" rel="noopener"><strong>Extension API</strong><span>Bring your model or SAE β</span></a><a href="https://github.com/OpenSAE/SAEBase/blob/main/docs/first-experiment.md" target="_blank" rel="noopener"><strong>First experiment</strong><span>Step-by-step tutorial β</span></a></div></section> |
| <footer class="footer"><span>OpenSAE / SAEBase Β· Evidence-backed interpretability infrastructure</span><span><a href="https://github.com/OpenSAE/SAEBase" target="_blank" rel="noopener">GitHub</a> Β· <a href="https://huggingface.co/OpenSAE" target="_blank" rel="noopener">OpenSAE on HF</a></span></footer> |
| </main> |
| </div> |
| <script src="app.js" defer></script> |
| </body> |
| </html> |
|
|
